alembic_fumes

> ripping off gerrymandle.com

I'm used to game developers sometimes displaying their political preferences through their art, but I have to say that Gerrymandle's "Win the election for orange." is still quite on the nose!

Seems like a reasonable policy to me, depending on the sector at the least? Even with open weight / self-hosted models, there's no way to know if the model has been fine-tuned to contain malicious behaviour, triggered only within very specific conditions.

In fairness this doesn't warrant the way that the policy targets Chinese models specifically, but Chinese providers do fall under the broader category of entities that cannot be prosecuted by the U.S. law if such behaviour is found.

So on one hand the page says

> We want to make dealing with agentic infrastructure easier so you can focus on your work. AX is designed with an uncompromising focus on ergonomics, rapid iteration, and joyful workflows for both application developers and AI researchers.

On the the other hand, the readme quickstart section says

> You need a Kubernetes cluster, ko (brew install ko), a container registry your cluster can pull from, and a reachable Agent Substrate Control API (in-cluster default: api.ate-system.svc.cluster.local:443).

Call me old-fashioned but I don't find this "easier". Maybe it's easier in the same way that Kubernetes itself is easier than managing VMs and container deployments at massive scale without such a tool. But there's a vast chasm between what this tool is being sold as and what it actually is.

I am having a hard time believing that anyone who uses the word "bro" this much is capable of advising anyone with seniority, because I don't see how they themselves would have the necessary seniority to give such advise.

But I suppose the message, hard to read as it is, is something that some people need to hear. I've seen this first hand a few times, mostly in people who got promoted by working really hard and the more they got promoted the harder they felt like they must work to keep proving themselves. This often led into burnout.

I don't know; I don't really want to blame the people themselves or say that they're the ones responsible of their own condition. People, and I want to say technically minded people in particular, are each rather fallible when it comes to our own humanity and the psychology underneath. And so, I'd much rather see this same advise be targeted to managers and senior teammates: when you notice a colleague who seems to have gone in too deep, have a friendly chat. They might very well be blind to it, or trying to ignore it to push through just that one more time.

This is a very interesting benchmark, and I think it has a lot of potential to make the speed of a model quantifiable.

I'm often asking myself is it better to use higher or lower effort levels, or to maybe drop down to a "dumber" but faster model. And so using a real-time based competition as a benchmark could shed some light on this, I think.

In this vein, here are what I would love to see added in this benchmark:

- Include Google's Gemini models. I keep hearing Gemini being praised for its speed, and I would like to see whether that gives it a big enough edge over the bigger but slower models. - How does a Cerebras-accelerated open source model fare against a much larger but much slower frontier model?

I also feel like in general there is a lot of very low-hanging fruit to start benchmarking models across the spectrum of real-time vs batch-style workloads. Perhaps Brood War sits somewhere quite near the "real-time" end of the spectrum, but what about something like a game of speed chess, or a turn-based game with time limits?

I think what I would like to see the most is for someone to come up with a benchmark that supports tuning the "real-timeliness" of the benchmark, and then running a sweep of a model across the whole spectrum. That could get result in real nice graphs with multiple models on the pareto-frontier, varying based on the hosting provider and the model dimensions.

I relate to this.

For years I used Rust as my hobby-programming language and loved it greatly. I never managed to land a job working with it full-time, because either the work was too niche or it didn't pay enough, or I was simply too comfortable where I was to change. And now that I finally have enough discretion over my technology choices to run a "proper" project using whatever tools and languages I want, it is not me but the AI that writes all of the code.

There's a part of me that feels a quite sad about all this. It's almost as if there actually all along existed a real final deadline on finding that "dream job". And I missed it. And while I expect this one miss to be just a small piece in the grand picture of things that we're going to lose or have already lost to the zeitgeist of agentic SWE, it feels big to me. It was my professional dream, while I still had professional dreams.

I did somewhen toy with the idea of comparing how a line of code is written to what an LLM-based next-token-prediction would emit for the same line. Then it might be possible to highlight the pieces of code that most differ from what an LLM would have written.

The base form isn't too useful by itself, but I expect if one would "seed" the LLM with a prompt about focusing specifically on a particular class of errors, it might be possible to find violations by seeing where the code completely deviates from what an LLM would have written.

Never got around to doing anything with this, but I expect a prototype should take no more than a day with some agent-assisted coding ;) Tell me how it went if you try!

> Someone has to put in the effort to make the markets accurate, and that someone has to be paid and that money has to come from somewhere.

The foundational idea of prediction markets is that this payment comes from the market itself. If you have a market full of suckers, the experts with real knowledge are incentivized to participate in the market to profit from the gap between what the market forecasts and their own expert forecast. This in turn will drive the market's prediction to be more accurate by incorporating the expert knowledge in directly form of the their "bets". In effect the market says put up or shut up to everyone who thinks they know better the market.

Viewed like this, prediction markets aren't much different from stock markets that also work by the premise of (active) investors claiming to know better than the market. It all follows from the efficient-market hypothesis.

All this said, I find the trend of prediction markets being used to predict what are effectively mostly random outcomes seems a bit strange. But on the other hand, the mere existence of such markets does provide financial incentive to come up with new, better ways to predict these outcomes. This itself would be very useful, at least as long as the subject is more akin to predicting the weather than the movements of a football.

Ultimately, I'm still an optimist when it comes to prediction markets.

> This is the true horror of the cognitive dark forest: it doesn’t kill you. It lets you live and feeds on you. Your innovation becomes its capabilities. Your differentiation becomes its median.

Oh no, the terrible dystopia where anyone can benefit from anyone else's good ideas without restrictions! And without any gatekeepers, licensing agreements, copyright, and not even a lawyer in sight!

If this is the dark future that AI use brings for us, I say bring it. Even if it means that somebody gets filthy rich in the process, while making the rest of the humanity better off.

The author almost touches on the one more topic that I adore about Nix, but ends up just so missing it: NixOS is absolutely incredible for its ability to be configured through AI tooling. And I don't mean that it's better than other operating systems, I mean that it's the only game in town.

I've been using Nix, both the package manager and the operating system, for years by now. I agree with all of the author's points, it really does deliver, the declarative nature is superb, and there's this constant sense of "hey my stuff is not breaking by itself" when working on it. And it's that declarative, rollback-able, file-based foundation, that makes it the perfect operating system for telling a coding agent to go to town on.

Would I trust Claude to switch my audio stack from Pulseaudio to Pipewire on Ubuntu? Would I trust Codex to install Hyprland on Fedora so I can test out the session? No, in fact I would not trust any agent to do any of those things on any other operating system. But I would trust even goddamn Grok to do that on NixOS, because I can 1) audit the changes before anything is done, and 2) rollback, rollforward, roll-whatever-the-way-I-want-even-on-the-floor-if-I-want-to because of the years of built up confidence proving that IT JUST WORKS.

I concede that this is turning into an unhinged loveletter to Nix, but really, it's the only operating system that lets one operate with this level of confidence. And I know most people don't care about that, since most people don't usually bother to tweak their OSes or switch out window managers, but as someone that does that, I'm never going back to mutable distros. This security is my table-stakes now, and the others aren't willing to pay up.

So for the developers out there on the lookout for their "Year of the Linux Desktop 2026" -distribution, if you're already using AI assistants, give NixOS a try. Maybe start with this in an empty Git repository: "Hey Claude, I wanna try NixOS. Make me a Flake-based starter config using Gnome that I can demo in a virtual machine. If nix isn't yet installed, install it via determinate-systems installer. Include a "vm" target in the flake for building the image, and a small bash script that builds and launches the VM using whatever virtualization is available on my platform."

> Software enables the enforcement of arbitrary rules that no human being would have the heart or foolishness to enforce.

For a long time I've been wanting to write an essay about this exact topic. Once you start paying attention, it's everywhere: the automatic upselling attempts at self-service kiosks, the automatic pre-filled tipping inputs in the card payment terminals, and some horribly long forms that require extremely pedantic input filled just correctly, that no human would bother on paper.

I've come to think of this as a kind of automated tyranny. Few real humans would have the heart to keep trying to upsell with such persistence when most of their customers reject the sale. And very few humans would dare to ask every single human to tip a pre-filled amount of 10% or even up in such a bold manner. But because "the machine does it", there's no alternative but to accept it, or to boycott the machines.

My condolences to the author for having to witness their children live through such an experience. It sounds like hell.

> strict is now true by default

I would still have a full head of hair if this had been the case since the beginning. Nonetheless I am glad that we got here in the end.

> Over a thousand pull requests merged each week at Stripe are completely minion-produced, and while they’re human-reviewed, they contain no human-written code.

I pity the senior engineer, demoted from a helmsman into a human breakwater, tasked to stand steady against an ever-swelling sea of AI slop.

I invite you to ponder the question: would a worldwide ban of all advertising have a greater or smaller impact on environment-destroying activity than banning of all air travel?

I would argue for "greater", and from that it rather naturally follows that advertisement and marketing indeed kills humanity.